From a8529ce2bed339b6f2200c7763f8007f4f79fb23 Mon Sep 17 00:00:00 2001 From: Yassin Soliman Date: Tue, 20 Oct 2026 11:48:35 -0600 Subject: [PATCH] Document the headers #3773 removed, and keep the one that is a pure rename The CK develop import dropped 13 headers that composablekernel-devel shipped in ROCm 7.0.2.1 compute build 50, including the whole ck::tensor_operation::device StreamK GEMM API. Measured on the published RPM file lists for builds 50 and 52: all 13 present at 50, none at 52, while the package grew from 1,298 to 1,690 files. Nothing on this branch records that, so a consumer compiling against the installed headers finds out at compile time after a patch-level update. Add a 7.0.2.1 changelog section listing every removal with its replacement, and restore device_grouped_conv_bwd_weight_explicit_xdl.hpp as a thin compatibility header. That one is a pure rename to device_grouped_conv_bwd_weight_explicit.hpp with an identical template parameter list, so the old spelling is an alias. --- CHANGELOG.md | 13 +++++ ...e_grouped_conv_bwd_weight_explicit_xdl.hpp | 47 +++++++++++++++++++ 2 files changed, 60 insertions(+) create mode 100644 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit_xdl.hpp diff --git a/CHANGELOG.md b/CHANGELOG.md index af318904f6..02c582527f 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -2,6 +2,19 @@ Documentation for Composable Kernel available at [https://rocm.docs.amd.com/projects/composable_kernel/en/latest/](https://rocm.docs.amd.com/projects/composable_kernel/en/latest/). +## Composable Kernel 1.2.0 for ROCm 7.0.2.1 + +### Removed + +The CK develop import (#3773) dropped 13 headers that `composablekernel-devel` shipped in ROCm 7.0.2.1 compute build 50. They are listed here with their replacements, measured on the published RPM file lists for builds 50 and 52. + +* Removed the `ck::tensor_operation::device` StreamK GEMM API: `ck/tensor_operation/gpu/device/device_gemm_streamk.hpp`, `device_gemm_streamk_v2.hpp`, `impl/device_gemm_xdl_cshuffle_streamk_v3.hpp`, `impl/device_gemm_xdl_streamk.hpp`, `grid/gridwise_gemm_xdl_cshuffle_streamk_v3.hpp`, `grid/gridwise_gemm_xdlops_streamk.hpp`, and the instance headers `ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp` and `gemm_universal_streamk.hpp`. The matching instance sources, examples and profiler op were removed with them. Use the CK Tile StreamK GEMM under `ck_tile/ops/gemm/kernel/streamk_gemm/`, which the same import added. +* Removed `ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp`. Use `naive_conv_fwd_gpu.hpp`, which is unchanged from build 50. +* Removed `ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp`. Use `device_grouped_conv_bwd_weight_v3_wmma_instance.hpp`, or the new `device_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp` and `device_grouped_conv_bwd_weight_wmma_scale_instance.hpp`. +* Removed `ck_tile/ops/fmha/pipeline/block_fmha_bwd_pipeline_enum.hpp`. `BlockFmhaBwdPipelineEnum` no longer exists and has no replacement. +* Renamed `ck_tile/ops/reduce/block/block_reduce2d_default_policy.hpp` to `ck_tile/ops/reduce/pipeline/reduce2d_default_policy.hpp`, and `BlockReduce2dDefaultPolicy` to `Reduce2dDefaultPolicy`. The policy now reads `kOutputIndex` from the problem and `ThreadTile_M`/`ThreadTile_N` from the block shape, so this is a source change for callers, not just an include path change. +* Renamed `ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit_xdl.hpp` to `device_grouped_conv_bwd_weight_explicit.hpp`, and `DeviceGroupedConvBwdWeight_Explicit_Xdl` to `DeviceGroupedConvBwdWeight_Explicit`. The old header and the old class name stay available for the 7.0.2.x series through a compatibility header; the template parameter list is unchanged. + ## Composable Kernel 1.2.0 for ROCm 7.13 ### Added diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit_xdl.hpp new file mode 100644 index 0000000000..26b0d9d747 --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit_xdl.hpp @@ -0,0 +1,47 @@ +// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. +// SPDX-License-Identifier: MIT + +#pragma once + +// Compatibility header for the ROCm 7.0.2.x series. +// +// composablekernel-devel shipped this path in ROCm 7.0.2.1 compute build 50. +// The CK develop import (#3773) renamed the file to +// device_grouped_conv_bwd_weight_explicit.hpp and the class from +// DeviceGroupedConvBwdWeight_Explicit_Xdl to DeviceGroupedConvBwdWeight_Explicit. +// The template parameter list is identical, so the old spelling stays available +// here for the rest of 7.0.2.x. New code should use the new path and name. + +#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_explicit.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +template +using DeviceGroupedConvBwdWeight_Explicit_Xdl = + DeviceGroupedConvBwdWeight_Explicit; + +} // namespace device +} // namespace tensor_operation +} // namespace ck